Overview
PeachDesk Voice Studio records and transcribes audio in a few different contexts, each feeding a different surface in the app. Everything runs on Bibha’s hosted GPU infrastructure — there’s nothing to install and no models to download. This page is a map; follow the links for the detail. Open Voice Studio from your PeachDesk dashboard to get started.
All paths share the same speech-to-text backend running on PeachDesk’s servers — it’s the surrounding workflow that differs.
Captures tab
When you want a clean transcript of some audio, the Captures tab is home. Record directly in your browser, or upload an existing file (.wav, .mp3, .m4a, .webm, .opus, .flac). Every capture keeps its original audio, can be re-transcribed with a different model, and can be played back through any voice profile you have.
The transcription model and decoding controls are covered under Transcription models below.
Dictation
When you want to turn speech into text quickly, record in the Captures tab and let Voice Studio transcribe it for you. The transcript is cleaned up automatically if auto-refine is enabled, and the audio is archived alongside it so you can replay or re-transcribe later. The live dictation panel is covered in Talk.Voice profile samples
A separate flow lives in the Voices tab. When you’re creating a profile with Clone from audio, the sample is what the cloning engine actually learns from — the reference text on a sample must match the audio verbatim, which is why samples are a different data model from captures. You can promote a capture to a sample from the Captures tab’s send-to menu (“Use as voice sample…”), which opens a reference-text confirm dialog so you can correct the last ~10% of transcript accuracy before saving. Covered in Voice Cloning.Transcription models
All paths share the same hosted speech-to-text family — LIPI, by Bibha AI Labs — served on PeachDesk’s GPU infrastructure. You never download or manage model files.
Choose a transcription model in Settings → Captures → Transcription (the list shows what your plane currently serves) or per capture — and re-transcribe a capture with a different one whenever you want to compare. The quality/latency trade-off within a model is a separate beam size control: Fast (greedy), Balanced, or Accurate (slow).
Because models run on PeachDesk’s servers, you don’t choose or configure a hardware backend — the platform routes your request to the right infrastructure automatically.
On very noisy or near-silent clips any speech-to-text model can hallucinate — most famously the “thanks for watching” loop. Voice Studio strips those loops deterministically before refinement runs, so a capture can be cleanly re-refined even if the raw transcript contains them. Enabling the voice-activity pre-filter, which trims silence before transcription, is the main defence.
Language
You can pass a language hint for short clips (under ~5 seconds) where automatic language detection is unreliable. Set a default language lock in Settings → Captures → Transcription → Language, or override it per capture.Next steps
Talk
Dictate, refine, and review the paired audio + transcript archive.
Voice Cloning
Record or upload samples for voice cloning.
Generating Speech
Turn text into natural-sounding speech.